A historical pattern:
1. A few brave individuals question authority and status quo: social, ideological, and philosophical.
2. These brave few begin to explore.
3. A new discovery, method, opportunity, or terrain is uncovered.
4. Authorities and the Fearful rush in to draw boundaries. Often these boundaries amputate important aspects of the discovery, especially open-endedness.
5. Dogma forms within these boundaries.
6. Students learn the dogma from the authorities.
7. Teachers omit the mindset required for exploration.
8. Dogma kills exploration for everybody.
9. Progress languishes.
There are so many examples of this. Off the top of my head, it reminds me of Peter Thiel's book Zero to One. Zero-to-one requires boldly exploring (step 1-3), then one-to-many kills exploration (steps 4-9).
It should not be debated that dogma be rejected in the sciences. There is no place for it. But what about sneaky dogma? Dogma that is so culturally dominant that it's invisible as a position: the formalization of mathematics.
"None of the 'creative' periods and hardly any of the 'critical' periods of mathematical theories would be admitted into the formalist heaven, where mathematical theories dwell like the seraphim, purged of all the impurities of earthly uncertainty."
- Imre Lakatos, Proofs and Refutations [1976], p. 2
Lakatos recounts the origins of mathematical "formal systems" and "abbreviatory devices" and what was lost in the process, namely, exploration in the form of "guesses by speculation and criticism." (quotes by Lakatos, Proofs and Refutations [1976])
Two modern textbook examples:
- How to Prove It by Daniel J. Velleman
- Book of Proof by Richard Hammack.
The books are wonderful, useful, and fun! The issue is believing that truth lives inside this formalization boundary because it does not.
Dogma feels safe. Yet, dogma is a disease to creativity, exploration, and progress. Exploration and explanatory space are infinite, have a little courage.
@sebkrier I agree. One doesn’t get alignment for free, but it’s seems to me that collaboration is a more beneficial and default position for any mind, at any capability tier. I’ve gone through pains to spell this out the my essay structural convergence.
“Everyone undervalues…”
“People aren’t ready for…”
“Nobody is talking about…”
X is filled with these hot takes, but what do they mean? It seems to me the scarce resource is attention.
Finding what is insightful, valuable, and applicable is a valuable skill. But, hey, I’m sure this is lost in the noise…
“Everyone undervalues…”
“People aren’t ready for…”
“Nobody is talking about…”
X is filled with these hot takes, but what do they mean? It seems to me the scarce resource is attention.
Finding what is insightful, valuable, and applicable is a valuable skill. But, hey, I’m sure this is lost in the noise…
I’ve certainly written about it. The implications may be stranger than even this post shows. My blog has a post for the computation conjecture that dives deeper. In my experience, the overwhelming majority of people have a “what is now” rubric of reasoning and are allergic to these topics. It is the pioneers, visionaries, and creatives who think along these lines. Living on the edge means seeing the future while being quite alone, save maybe a small handful of close friends. The people you speak of will talk about it when it is in their proverbial face.
I like to think about the abstract boundary you describe as the interface layer (of abstraction.) Same thing, different words.
For example, LLMs provide a new solution interface (not UI) with new benefits and costs. Interesting properties emerge at each interface layer because one loses some capabilities while gaining others. A technology’s impact is the difference between the old and new interface layer along costs/benefits axes. What I find most intriguing are the benefits that relate to solving old problems and new capabilities.
Take software engineering, no need to write LOC, but now one must manage agents. The impact then is how competition impacts markets.
Just this week, my agentic system (open source) orchestrated complex software changes leveraging non-technical staff in a distributed fashion. New problems emerge, but the benefit vastly outweighs the costs. This is starting to have systemic impacts on the enterprise using this interface layer. More than solving old problems, it’s reasonable to expect the new capabilities will incentivize business into new (maybe latent) frontiers.
I am rambling a bit. Sitting in the ER it’s my toddler. He had an asthma attack. No fun, but he’s going to be okay.
I live above all models, but cherish them. Of course, I don’t expect this to compel anyone. That’s okay. For example, I’m sitting in the ER right now with my toddler who is on a breathing machine. I see the monitors, hear the beeps, and appreciate the quantification of human health. The math and machines are amazing. Knowing what a normal heart rate and oxygen level is could very well have saved my son’s life today. On the other hand, I also know that reality is infinite and we aren’t done exploring it, not by a long-shot. Living the between ungroundedness and dogmatism is where wisdom grows.
Thanks! One important note, I don't use the second agent for surface verification. I have another app/repo with deterministic tests for that. So the split agents are really about implementation, not really about outcome control, "Hey, I asked agent A to do a thing. Mr. Agent B, did it really do the thing?"
Codex + Claude Code did for me in one day what I could not have done months: reimplement my entire project in rust. Codex implemented and Claude verified. I started at around 5pm writing a spec, started the run around 7am and about 12 hours later the work is done. Incredible!
The project is called Refine. It is decentralized agentic coding - leveraging the very tools I used today - at scale. It is designed to coordinate teams of people, across many nodes on legacy/new code bases without new infrastructure. I wish you knew how big of a deal that is.
It's an achievement to you have a long agentic session in the background, assuming you get something great at the other end:
Goal usage: 6,629,886 tokens over about 6h 14m.
Excited! Refine is coming to the desktop. Under the hood, refine is being ported to rust. Rust has attractive properties that make it an easy architectural decision long term. The goal is to use AI Personal Assistant (Claude Cowork, Codex, OpenClaw) to manage a cluster of agents on the beach/mountains/space.
It’s even worse than that.
Consider: math does not generate theories, instead math is a tool used to aid the transition from theory to practice and to criticize theories. The brain does not run on math, but can be quantified using a mathematical perspective. Math is about quantities. That is why there is a whole category called qualities.
Math is a model and with all models incomplete. Obviously, those who dedicate their whole lives to it see it otherwise, also known as the Tetris effect.
So, I support your position. It is common to see dogmatic mathematicians, but it is common to see dogmatists in every discipline.
I have written about both the topics of models/math and fundamental ontologies in my blog: Living Above the Models and Computation Conjecture.
AI coding agents are great at the first 80%, but that last 20% will consume a lot of time. Agents are great, but state is a hard problem.
Another reason to use Refine!
DeepMind, Anthropic, and OpenAI all backing medicine, robotics, and enterprise software - an obvious gold rush - big tech is exploding in progress all thanks to LLMs.
Leveraging existing infrastructure is a product decision I didn't see clearly until now. I see it everywhere now. I've made an impressive decentralized agentic software delivery platform: refine - if I do say so myself. It's essentially a wrapper around CLI agents. It useful because it coordinates people and agents on existing hardware, not a new cloud or centralized service.
Here's what I missed: the refine app is new - it doesn't leverage existing tools project management tools like Jira - so I'm not seeing a long runway for it. Now that agents can orchestrate 100s of agents, it won't be long before they can scale this orchestration to 100s of machines. Then they'll write some easy integrations with existing project management software. Once that exists, I wonder what value my project has? For now, it's the best in class, but for how long?
I'll continue to work on refine because it is needed by my enterprise clients, and maybe it'll be useful in the future, but I'm having a hard time seeing it right now.
The main way I get stuck in writer's block is when I haven't dealt with the strongest argument against my position head on as early as possible.
And because I haven't, I have to dance around with what I'm trying to say so as to not active the rebuttal.